lustre/lov/lov_request.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #ifndef EXPORT_SYMTAB
  38 # define EXPORT_SYMTAB
  39 #endif
  40 #define DEBUG_SUBSYSTEM S_LOV
  41
  42 #ifdef __KERNEL__
  43 #include <libcfs/libcfs.h>
  44 #else
  45 #include <liblustre.h>
  46 #endif
  47
  48 #include <obd_class.h>
  49 #include <obd_lov.h>
  50 #include <lustre/lustre_idl.h>
  51
  52 #include "lov_internal.h"
  53
  54 static void lov_init_set(struct lov_request_set *set)
  55 {
  56         set->set_count = 0;
  57         set->set_completes = 0;
  58         set->set_success = 0;
  59         set->set_cookies = 0;
  60         CFS_INIT_LIST_HEAD(&set->set_list);
  61         atomic_set(&set->set_refcount, 1);
  62 }
  63
  64 static void lov_finish_set(struct lov_request_set *set)
  65 {
  66         struct list_head *pos, *n;
  67         ENTRY;
  68
  69         LASSERT(set);
  70         list_for_each_safe(pos, n, &set->set_list) {
  71                 struct lov_request *req = list_entry(pos, struct lov_request,
  72                                                      rq_link);
  73                 list_del_init(&req->rq_link);
  74
  75                 if (req->rq_oi.oi_oa)
  76                         OBDO_FREE(req->rq_oi.oi_oa);
  77                 if (req->rq_oi.oi_md)
  78                         OBD_FREE(req->rq_oi.oi_md, req->rq_buflen);
  79                 if (req->rq_oi.oi_osfs)
  80                         OBD_FREE(req->rq_oi.oi_osfs,
  81                                  sizeof(*req->rq_oi.oi_osfs));
  82                 OBD_FREE(req, sizeof(*req));
  83         }
  84
  85         if (set->set_pga) {
  86                 int len = set->set_oabufs * sizeof(*set->set_pga);
  87                 OBD_FREE(set->set_pga, len);
  88         }
  89         if (set->set_lockh)
  90                 lov_llh_put(set->set_lockh);
  91
  92         OBD_FREE(set, sizeof(*set));
  93         EXIT;
  94 }
  95
  96 void lov_update_set(struct lov_request_set *set,
  97                     struct lov_request *req, int rc)
  98 {
  99         req->rq_complete = 1;
 100         req->rq_rc = rc;
 101
 102         set->set_completes++;
 103         if (rc == 0)
 104                 set->set_success++;
 105 }
 106
 107 int lov_update_common_set(struct lov_request_set *set,
 108                           struct lov_request *req, int rc)
 109 {
 110         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 111         ENTRY;
 112
 113         lov_update_set(set, req, rc);
 114
 115         /* grace error on inactive ost */
 116         if (rc && !(lov->lov_tgts[req->rq_idx] &&
 117                     lov->lov_tgts[req->rq_idx]->ltd_active))
 118                 rc = 0;
 119
 120         /* FIXME in raid1 regime, should return 0 */
 121         RETURN(rc);
 122 }
 123
 124 void lov_set_add_req(struct lov_request *req, struct lov_request_set *set)
 125 {
 126         list_add_tail(&req->rq_link, &set->set_list);
 127         set->set_count++;
 128 }
 129
 130 extern void osc_update_enqueue(struct lustre_handle *lov_lockhp,
 131                                struct lov_oinfo *loi, int flags,
 132                                struct ost_lvb *lvb, __u32 mode, int rc);
 133
 134 static int lov_update_enqueue_lov(struct obd_export *exp,
 135                                   struct lustre_handle *lov_lockhp,
 136                                   struct lov_oinfo *loi, int flags, int idx,
 137                                   __u64 oid, int rc)
 138 {
 139         struct lov_obd *lov = &exp->exp_obd->u.lov;
 140
 141         if (rc != ELDLM_OK &&
 142             !(rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT))) {
 143                 memset(lov_lockhp, 0, sizeof(*lov_lockhp));
 144                 if (lov->lov_tgts[idx] && lov->lov_tgts[idx]->ltd_active) {
 145                         /* -EUSERS used by OST to report file contention */
 146                         if (rc != -EINTR && rc != -EUSERS)
 147                                 CERROR("enqueue objid "LPX64" subobj "
 148                                        LPX64" on OST idx %d: rc %d\n",
 149                                        oid, loi->loi_id, loi->loi_ost_idx, rc);
 150                 } else
 151                         rc = ELDLM_OK;
 152         }
 153         return rc;
 154 }
 155
 156 int lov_update_enqueue_set(struct lov_request *req, __u32 mode, int rc)
 157 {
 158         struct lov_request_set *set = req->rq_rqset;
 159         struct lustre_handle *lov_lockhp;
 160         struct obd_info *oi = set->set_oi;
 161         struct lov_oinfo *loi;
 162         ENTRY;
 163
 164         LASSERT(oi != NULL);
 165
 166         lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe;
 167         loi = oi->oi_md->lsm_oinfo[req->rq_stripe];
 168
 169         /* XXX LOV STACKING: OSC gets a copy, created in lov_prep_enqueue_set
 170          * and that copy can be arbitrarily out of date.
 171          *
 172          * The LOV API is due for a serious rewriting anyways, and this
 173          * can be addressed then. */
 174
 175         lov_stripe_lock(oi->oi_md);
 176         osc_update_enqueue(lov_lockhp, loi, oi->oi_flags,
 177                            &req->rq_oi.oi_md->lsm_oinfo[0]->loi_lvb, mode, rc);
 178         if (rc == ELDLM_LOCK_ABORTED && (oi->oi_flags & LDLM_FL_HAS_INTENT))
 179                 memset(lov_lockhp, 0, sizeof *lov_lockhp);
 180         rc = lov_update_enqueue_lov(set->set_exp, lov_lockhp, loi, oi->oi_flags,
 181                                     req->rq_idx, oi->oi_md->lsm_object_id, rc);
 182         lov_stripe_unlock(oi->oi_md);
 183         lov_update_set(set, req, rc);
 184         RETURN(rc);
 185 }
 186
 187 /* The callback for osc_enqueue that updates lov info for every OSC request. */
 188 static int cb_update_enqueue(void *cookie, int rc)
 189 {
 190         struct obd_info *oinfo = cookie;
 191         struct ldlm_enqueue_info *einfo;
 192         struct lov_request *lovreq;
 193
 194         lovreq = container_of(oinfo, struct lov_request, rq_oi);
 195         einfo = lovreq->rq_rqset->set_ei;
 196         return lov_update_enqueue_set(lovreq, einfo->ei_mode, rc);
 197 }
 198
 199 static int enqueue_done(struct lov_request_set *set, __u32 mode)
 200 {
 201         struct lov_request *req;
 202         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 203         int rc = 0;
 204         ENTRY;
 205
 206         /* enqueue/match success, just return */
 207         if (set->set_completes && set->set_completes == set->set_success)
 208                 RETURN(0);
 209
 210         /* cancel enqueued/matched locks */
 211         list_for_each_entry(req, &set->set_list, rq_link) {
 212                 struct lustre_handle *lov_lockhp;
 213
 214                 if (!req->rq_complete || req->rq_rc)
 215                         continue;
 216
 217                 lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe;
 218                 LASSERT(lov_lockhp);
 219                 if (!lustre_handle_is_used(lov_lockhp))
 220                         continue;
 221
 222                 rc = obd_cancel(lov->lov_tgts[req->rq_idx]->ltd_exp,
 223                                 req->rq_oi.oi_md, mode, lov_lockhp);
 224                 if (rc && lov->lov_tgts[req->rq_idx] &&
 225                     lov->lov_tgts[req->rq_idx]->ltd_active)
 226                         CERROR("cancelling obdjid "LPX64" on OST "
 227                                "idx %d error: rc = %d\n",
 228                                req->rq_oi.oi_md->lsm_object_id,
 229                                req->rq_idx, rc);
 230         }
 231         if (set->set_lockh)
 232                 lov_llh_put(set->set_lockh);
 233         RETURN(rc);
 234 }
 235
 236 int lov_fini_enqueue_set(struct lov_request_set *set, __u32 mode, int rc,
 237                          struct ptlrpc_request_set *rqset)
 238 {
 239         int ret = 0;
 240         ENTRY;
 241
 242         if (set == NULL)
 243                 RETURN(0);
 244         LASSERT(set->set_exp);
 245         /* Do enqueue_done only for sync requests and if any request
 246          * succeeded. */
 247         if (!rqset) {
 248                 if (rc)
 249                         set->set_completes = 0;
 250                 ret = enqueue_done(set, mode);
 251         } else if (set->set_lockh)
 252                 lov_llh_put(set->set_lockh);
 253
 254         if (atomic_dec_and_test(&set->set_refcount))
 255                 lov_finish_set(set);
 256
 257         RETURN(rc ? rc : ret);
 258 }
 259
 260 int lov_prep_enqueue_set(struct obd_export *exp, struct obd_info *oinfo,
 261                          struct ldlm_enqueue_info *einfo,
 262                          struct lov_request_set **reqset)
 263 {
 264         struct lov_obd *lov = &exp->exp_obd->u.lov;
 265         struct lov_request_set *set;
 266         int i, rc = 0;
 267         ENTRY;
 268
 269         OBD_ALLOC(set, sizeof(*set));
 270         if (set == NULL)
 271                 RETURN(-ENOMEM);
 272         lov_init_set(set);
 273
 274         set->set_exp = exp;
 275         set->set_oi = oinfo;
 276         set->set_ei = einfo;
 277         set->set_lockh = lov_llh_new(oinfo->oi_md);
 278         if (set->set_lockh == NULL)
 279                 GOTO(out_set, rc = -ENOMEM);
 280         oinfo->oi_lockh->cookie = set->set_lockh->llh_handle.h_cookie;
 281
 282         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
 283                 struct lov_oinfo *loi;
 284                 struct lov_request *req;
 285                 obd_off start, end;
 286
 287                 loi = oinfo->oi_md->lsm_oinfo[i];
 288                 if (!lov_stripe_intersects(oinfo->oi_md, i,
 289                                            oinfo->oi_policy.l_extent.start,
 290                                            oinfo->oi_policy.l_extent.end,
 291                                            &start, &end))
 292                         continue;
 293
 294                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
 295                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
 296                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 297                         continue;
 298                 }
 299
 300                 OBD_ALLOC(req, sizeof(*req));
 301                 if (req == NULL)
 302                         GOTO(out_set, rc = -ENOMEM);
 303
 304                 req->rq_buflen = sizeof(*req->rq_oi.oi_md) +
 305                         sizeof(struct lov_oinfo *) +
 306                         sizeof(struct lov_oinfo);
 307                 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
 308                 if (req->rq_oi.oi_md == NULL) {
 309                         OBD_FREE(req, sizeof(*req));
 310                         GOTO(out_set, rc = -ENOMEM);
 311                 }
 312                 req->rq_oi.oi_md->lsm_oinfo[0] =
 313                         ((void *)req->rq_oi.oi_md) + sizeof(*req->rq_oi.oi_md) +
 314                         sizeof(struct lov_oinfo *);
 315
 316
 317                 req->rq_rqset = set;
 318                 /* Set lov request specific parameters. */
 319                 req->rq_oi.oi_lockh = set->set_lockh->llh_handles + i;
 320                 req->rq_oi.oi_cb_up = cb_update_enqueue;
 321                 req->rq_oi.oi_flags = oinfo->oi_flags;
 322
 323                 LASSERT(req->rq_oi.oi_lockh);
 324
 325                 req->rq_oi.oi_policy.l_extent.gid =
 326                         oinfo->oi_policy.l_extent.gid;
 327                 req->rq_oi.oi_policy.l_extent.start = start;
 328                 req->rq_oi.oi_policy.l_extent.end = end;
 329
 330                 req->rq_idx = loi->loi_ost_idx;
 331                 req->rq_stripe = i;
 332
 333                 /* XXX LOV STACKING: submd should be from the subobj */
 334                 req->rq_oi.oi_md->lsm_object_id = loi->loi_id;
 335                 req->rq_oi.oi_md->lsm_object_gr = oinfo->oi_md->lsm_object_gr;
 336                 req->rq_oi.oi_md->lsm_stripe_count = 0;
 337                 req->rq_oi.oi_md->lsm_oinfo[0]->loi_kms_valid =
 338                         loi->loi_kms_valid;
 339                 req->rq_oi.oi_md->lsm_oinfo[0]->loi_kms = loi->loi_kms;
 340                 req->rq_oi.oi_md->lsm_oinfo[0]->loi_lvb = loi->loi_lvb;
 341
 342                 lov_set_add_req(req, set);
 343         }
 344         if (!set->set_count)
 345                 GOTO(out_set, rc = -EIO);
 346         *reqset = set;
 347         RETURN(0);
 348 out_set:
 349         lov_fini_enqueue_set(set, einfo->ei_mode, rc, NULL);
 350         RETURN(rc);
 351 }
 352
 353 int lov_update_match_set(struct lov_request_set *set, struct lov_request *req,
 354                          int rc)
 355 {
 356         int ret = rc;
 357         ENTRY;
 358
 359         if (rc > 0)
 360                 ret = 0;
 361         else if (rc == 0)
 362                 ret = 1;
 363         lov_update_set(set, req, ret);
 364         RETURN(rc);
 365 }
 366
 367 int lov_fini_match_set(struct lov_request_set *set, __u32 mode, int flags)
 368 {
 369         int rc = 0;
 370         ENTRY;
 371
 372         if (set == NULL)
 373                 RETURN(0);
 374         LASSERT(set->set_exp);
 375         rc = enqueue_done(set, mode);
 376         if ((set->set_count == set->set_success) &&
 377             (flags & LDLM_FL_TEST_LOCK))
 378                 lov_llh_put(set->set_lockh);
 379
 380         if (atomic_dec_and_test(&set->set_refcount))
 381                 lov_finish_set(set);
 382
 383         RETURN(rc);
 384 }
 385
 386 int lov_prep_match_set(struct obd_export *exp, struct obd_info *oinfo,
 387                        struct lov_stripe_md *lsm, ldlm_policy_data_t *policy,
 388                        __u32 mode, struct lustre_handle *lockh,
 389                        struct lov_request_set **reqset)
 390 {
 391         struct lov_obd *lov = &exp->exp_obd->u.lov;
 392         struct lov_request_set *set;
 393         int i, rc = 0;
 394         ENTRY;
 395
 396         OBD_ALLOC(set, sizeof(*set));
 397         if (set == NULL)
 398                 RETURN(-ENOMEM);
 399         lov_init_set(set);
 400
 401         set->set_exp = exp;
 402         set->set_oi = oinfo;
 403         set->set_oi->oi_md = lsm;
 404         set->set_lockh = lov_llh_new(lsm);
 405         if (set->set_lockh == NULL)
 406                 GOTO(out_set, rc = -ENOMEM);
 407         lockh->cookie = set->set_lockh->llh_handle.h_cookie;
 408
 409         for (i = 0; i < lsm->lsm_stripe_count; i++){
 410                 struct lov_oinfo *loi;
 411                 struct lov_request *req;
 412                 obd_off start, end;
 413
 414                 loi = lsm->lsm_oinfo[i];
 415                 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
 416                                            policy->l_extent.end, &start, &end))
 417                         continue;
 418
 419                 /* FIXME raid1 should grace this error */
 420                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
 421                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
 422                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 423                         GOTO(out_set, rc = -EIO);
 424                 }
 425
 426                 OBD_ALLOC(req, sizeof(*req));
 427                 if (req == NULL)
 428                         GOTO(out_set, rc = -ENOMEM);
 429
 430                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 431                 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
 432                 if (req->rq_oi.oi_md == NULL) {
 433                         OBD_FREE(req, sizeof(*req));
 434                         GOTO(out_set, rc = -ENOMEM);
 435                 }
 436
 437                 req->rq_oi.oi_policy.l_extent.start = start;
 438                 req->rq_oi.oi_policy.l_extent.end = end;
 439                 req->rq_oi.oi_policy.l_extent.gid = policy->l_extent.gid;
 440
 441                 req->rq_idx = loi->loi_ost_idx;
 442                 req->rq_stripe = i;
 443
 444                 /* XXX LOV STACKING: submd should be from the subobj */
 445                 req->rq_oi.oi_md->lsm_object_id = loi->loi_id;
 446                 req->rq_oi.oi_md->lsm_object_gr = lsm->lsm_object_gr;
 447                 req->rq_oi.oi_md->lsm_stripe_count = 0;
 448
 449                 lov_set_add_req(req, set);
 450         }
 451         if (!set->set_count)
 452                 GOTO(out_set, rc = -EIO);
 453         *reqset = set;
 454         RETURN(rc);
 455 out_set:
 456         lov_fini_match_set(set, mode, 0);
 457         RETURN(rc);
 458 }
 459
 460 int lov_fini_cancel_set(struct lov_request_set *set)
 461 {
 462         int rc = 0;
 463         ENTRY;
 464
 465         if (set == NULL)
 466                 RETURN(0);
 467
 468         LASSERT(set->set_exp);
 469         if (set->set_lockh)
 470                 lov_llh_put(set->set_lockh);
 471
 472         if (atomic_dec_and_test(&set->set_refcount))
 473                 lov_finish_set(set);
 474
 475         RETURN(rc);
 476 }
 477
 478 int lov_prep_cancel_set(struct obd_export *exp, struct obd_info *oinfo,
 479                         struct lov_stripe_md *lsm, __u32 mode,
 480                         struct lustre_handle *lockh,
 481                         struct lov_request_set **reqset)
 482 {
 483         struct lov_request_set *set;
 484         int i, rc = 0;
 485         ENTRY;
 486
 487         OBD_ALLOC(set, sizeof(*set));
 488         if (set == NULL)
 489                 RETURN(-ENOMEM);
 490         lov_init_set(set);
 491
 492         set->set_exp = exp;
 493         set->set_oi = oinfo;
 494         set->set_oi->oi_md = lsm;
 495         set->set_lockh = lov_handle2llh(lockh);
 496         if (set->set_lockh == NULL) {
 497                 CERROR("LOV: invalid lov lock handle %p\n", lockh);
 498                 GOTO(out_set, rc = -EINVAL);
 499         }
 500         lockh->cookie = set->set_lockh->llh_handle.h_cookie;
 501
 502         for (i = 0; i < lsm->lsm_stripe_count; i++){
 503                 struct lov_request *req;
 504                 struct lustre_handle *lov_lockhp;
 505                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 506
 507                 lov_lockhp = set->set_lockh->llh_handles + i;
 508                 if (!lustre_handle_is_used(lov_lockhp)) {
 509                         CDEBUG(D_RPCTRACE,"lov idx %d subobj "LPX64" no lock\n",
 510                                loi->loi_ost_idx, loi->loi_id);
 511                         continue;
 512                 }
 513
 514                 OBD_ALLOC(req, sizeof(*req));
 515                 if (req == NULL)
 516                         GOTO(out_set, rc = -ENOMEM);
 517
 518                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 519                 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
 520                 if (req->rq_oi.oi_md == NULL) {
 521                         OBD_FREE(req, sizeof(*req));
 522                         GOTO(out_set, rc = -ENOMEM);
 523                 }
 524
 525                 req->rq_idx = loi->loi_ost_idx;
 526                 req->rq_stripe = i;
 527
 528                 /* XXX LOV STACKING: submd should be from the subobj */
 529                 req->rq_oi.oi_md->lsm_object_id = loi->loi_id;
 530                 req->rq_oi.oi_md->lsm_object_gr = lsm->lsm_object_gr;
 531                 req->rq_oi.oi_md->lsm_stripe_count = 0;
 532
 533                 lov_set_add_req(req, set);
 534         }
 535         if (!set->set_count)
 536                 GOTO(out_set, rc = -EIO);
 537         *reqset = set;
 538         RETURN(rc);
 539 out_set:
 540         lov_fini_cancel_set(set);
 541         RETURN(rc);
 542 }
 543
 544 static int create_done(struct obd_export *exp, struct lov_request_set *set,
 545                        struct lov_stripe_md **lsmp)
 546 {
 547         struct lov_obd *lov = &exp->exp_obd->u.lov;
 548         struct obd_trans_info *oti = set->set_oti;
 549         struct obdo *src_oa = set->set_oi->oi_oa;
 550         struct lov_request *req;
 551         struct obdo *ret_oa = NULL;
 552         int attrset = 0, rc = 0;
 553         ENTRY;
 554
 555         LASSERT(set->set_completes);
 556
 557         /* try alloc objects on other osts if osc_create fails for
 558          * exceptions: RPC failure, ENOSPC, etc */
 559         if (set->set_count != set->set_success) {
 560                 list_for_each_entry (req, &set->set_list, rq_link) {
 561                         if (req->rq_rc == 0)
 562                                 continue;
 563
 564                         set->set_completes--;
 565                         req->rq_complete = 0;
 566
 567                         rc = qos_remedy_create(set, req);
 568                         lov_update_create_set(set, req, rc);
 569
 570                         if (rc)
 571                                 break;
 572                 }
 573         }
 574
 575         /* no successful creates */
 576         if (set->set_success == 0)
 577                 GOTO(cleanup, rc);
 578
 579         /* If there was an explicit stripe set, fail.  Otherwise, we
 580          * got some objects and that's not bad. */
 581         if (set->set_count != set->set_success) {
 582                 if (*lsmp)
 583                         GOTO(cleanup, rc);
 584                 set->set_count = set->set_success;
 585                 qos_shrink_lsm(set);
 586         }
 587
 588         OBDO_ALLOC(ret_oa);
 589         if (ret_oa == NULL)
 590                 GOTO(cleanup, rc = -ENOMEM);
 591
 592         list_for_each_entry(req, &set->set_list, rq_link) {
 593                 if (!req->rq_complete || req->rq_rc)
 594                         continue;
 595                 lov_merge_attrs(ret_oa, req->rq_oi.oi_oa,
 596                                 req->rq_oi.oi_oa->o_valid, set->set_oi->oi_md,
 597                                 req->rq_stripe, &attrset);
 598         }
 599         if (src_oa->o_valid & OBD_MD_FLSIZE &&
 600             ret_oa->o_size != src_oa->o_size) {
 601                 CERROR("original size "LPU64" isn't new object size "LPU64"\n",
 602                        src_oa->o_size, ret_oa->o_size);
 603                 LBUG();
 604         }
 605         ret_oa->o_id = src_oa->o_id;
 606         ret_oa->o_gr = src_oa->o_gr;
 607         ret_oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
 608         memcpy(src_oa, ret_oa, sizeof(*src_oa));
 609         OBDO_FREE(ret_oa);
 610
 611         *lsmp = set->set_oi->oi_md;
 612         GOTO(done, rc = 0);
 613
 614 cleanup:
 615         list_for_each_entry(req, &set->set_list, rq_link) {
 616                 struct obd_export *sub_exp;
 617                 int err = 0;
 618
 619                 if (!req->rq_complete || req->rq_rc)
 620                         continue;
 621
 622                 sub_exp = lov->lov_tgts[req->rq_idx]->ltd_exp;
 623                 err = obd_destroy(sub_exp, req->rq_oi.oi_oa, NULL, oti, NULL);
 624                 if (err)
 625                         CERROR("Failed to uncreate objid "LPX64" subobj "
 626                                LPX64" on OST idx %d: rc = %d\n",
 627                                src_oa->o_id, req->rq_oi.oi_oa->o_id,
 628                                req->rq_idx, rc);
 629         }
 630         if (*lsmp == NULL)
 631                 obd_free_memmd(exp, &set->set_oi->oi_md);
 632 done:
 633         if (oti && set->set_cookies) {
 634                 oti->oti_logcookies = set->set_cookies;
 635                 if (!set->set_cookie_sent) {
 636                         oti_free_cookies(oti);
 637                         src_oa->o_valid &= ~OBD_MD_FLCOOKIE;
 638                 } else {
 639                         src_oa->o_valid |= OBD_MD_FLCOOKIE;
 640                 }
 641         }
 642         RETURN(rc);
 643 }
 644
 645 int lov_fini_create_set(struct lov_request_set *set,struct lov_stripe_md **lsmp)
 646 {
 647         int rc = 0;
 648         ENTRY;
 649
 650         if (set == NULL)
 651                 RETURN(0);
 652         LASSERT(set->set_exp);
 653         if (set->set_completes)
 654                 rc = create_done(set->set_exp, set, lsmp);
 655
 656         if (atomic_dec_and_test(&set->set_refcount))
 657                 lov_finish_set(set);
 658
 659         RETURN(rc);
 660 }
 661
 662 int lov_update_create_set(struct lov_request_set *set,
 663                           struct lov_request *req, int rc)
 664 {
 665         struct obd_trans_info *oti = set->set_oti;
 666         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 667         struct lov_oinfo *loi;
 668         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 669         ENTRY;
 670
 671         req->rq_stripe = set->set_success;
 672         loi = lsm->lsm_oinfo[req->rq_stripe];
 673
 674         if (rc && lov->lov_tgts[req->rq_idx] &&
 675             lov->lov_tgts[req->rq_idx]->ltd_active) {
 676                 CERROR("error creating fid "LPX64" sub-object"
 677                        " on OST idx %d/%d: rc = %d\n",
 678                        set->set_oi->oi_oa->o_id, req->rq_idx,
 679                        lsm->lsm_stripe_count, rc);
 680                 if (rc > 0) {
 681                         CERROR("obd_create returned invalid err %d\n", rc);
 682                         rc = -EIO;
 683                 }
 684         }
 685         lov_update_set(set, req, rc);
 686         if (rc)
 687                 RETURN(rc);
 688
 689         loi->loi_id = req->rq_oi.oi_oa->o_id;
 690         loi->loi_gr = req->rq_oi.oi_oa->o_gr;
 691         loi->loi_ost_idx = req->rq_idx;
 692         CDEBUG(D_INODE, "objid "LPX64" has subobj "LPX64"/"LPU64" at idx %d\n",
 693                lsm->lsm_object_id, loi->loi_id, loi->loi_id, req->rq_idx);
 694         loi_init(loi);
 695
 696         if (oti && set->set_cookies)
 697                 ++oti->oti_logcookies;
 698         if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLCOOKIE)
 699                 set->set_cookie_sent++;
 700
 701         RETURN(0);
 702 }
 703
 704 int lov_prep_create_set(struct obd_export *exp, struct obd_info *oinfo,
 705                         struct lov_stripe_md **lsmp, struct obdo *src_oa,
 706                         struct obd_trans_info *oti,
 707                         struct lov_request_set **reqset)
 708 {
 709         struct lov_request_set *set;
 710         int rc = 0;
 711         ENTRY;
 712
 713         OBD_ALLOC(set, sizeof(*set));
 714         if (set == NULL)
 715                 RETURN(-ENOMEM);
 716         lov_init_set(set);
 717
 718         set->set_exp = exp;
 719         set->set_oi = oinfo;
 720         set->set_oi->oi_md = *lsmp;
 721         set->set_oi->oi_oa = src_oa;
 722         set->set_oti = oti;
 723
 724         rc = qos_prep_create(exp, set);
 725         if (rc)
 726                 lov_fini_create_set(set, lsmp);
 727         else
 728                 *reqset = set;
 729         RETURN(rc);
 730 }
 731
 732 static int common_attr_done(struct lov_request_set *set)
 733 {
 734         struct list_head *pos;
 735         struct lov_request *req;
 736         struct obdo *tmp_oa;
 737         int rc = 0, attrset = 0;
 738         ENTRY;
 739
 740         LASSERT(set->set_oi != NULL);
 741
 742         if (set->set_oi->oi_oa == NULL)
 743                 RETURN(0);
 744
 745         if (!set->set_success)
 746                 RETURN(-EIO);
 747
 748         OBDO_ALLOC(tmp_oa);
 749         if (tmp_oa == NULL)
 750                 GOTO(out, rc = -ENOMEM);
 751
 752         list_for_each (pos, &set->set_list) {
 753                 req = list_entry(pos, struct lov_request, rq_link);
 754
 755                 if (!req->rq_complete || req->rq_rc)
 756                         continue;
 757                 if (req->rq_oi.oi_oa->o_valid == 0)   /* inactive stripe */
 758                         continue;
 759                 lov_merge_attrs(tmp_oa, req->rq_oi.oi_oa,
 760                                 req->rq_oi.oi_oa->o_valid,
 761                                 set->set_oi->oi_md, req->rq_stripe, &attrset);
 762         }
 763         if (!attrset) {
 764                 CERROR("No stripes had valid attrs\n");
 765                 rc = -EIO;
 766         }
 767         tmp_oa->o_id = set->set_oi->oi_oa->o_id;
 768         memcpy(set->set_oi->oi_oa, tmp_oa, sizeof(*set->set_oi->oi_oa));
 769 out:
 770         if (tmp_oa)
 771                 OBDO_FREE(tmp_oa);
 772         RETURN(rc);
 773
 774 }
 775
 776 static int brw_done(struct lov_request_set *set)
 777 {
 778         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 779         struct lov_oinfo     *loi = NULL;
 780         struct list_head *pos;
 781         struct lov_request *req;
 782         ENTRY;
 783
 784         list_for_each (pos, &set->set_list) {
 785                 req = list_entry(pos, struct lov_request, rq_link);
 786
 787                 if (!req->rq_complete || req->rq_rc)
 788                         continue;
 789
 790                 loi = lsm->lsm_oinfo[req->rq_stripe];
 791
 792                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLBLOCKS)
 793                         loi->loi_lvb.lvb_blocks = req->rq_oi.oi_oa->o_blocks;
 794         }
 795
 796         RETURN(0);
 797 }
 798
 799 int lov_fini_brw_set(struct lov_request_set *set)
 800 {
 801         int rc = 0;
 802         ENTRY;
 803
 804         if (set == NULL)
 805                 RETURN(0);
 806         LASSERT(set->set_exp);
 807         if (set->set_completes) {
 808                 rc = brw_done(set);
 809                 /* FIXME update qos data here */
 810         }
 811         if (atomic_dec_and_test(&set->set_refcount))
 812                 lov_finish_set(set);
 813
 814         RETURN(rc);
 815 }
 816
 817 int lov_prep_brw_set(struct obd_export *exp, struct obd_info *oinfo,
 818                      obd_count oa_bufs, struct brw_page *pga,
 819                      struct obd_trans_info *oti,
 820                      struct lov_request_set **reqset)
 821 {
 822         struct {
 823                 obd_count       index;
 824                 obd_count       count;
 825                 obd_count       off;
 826         } *info = NULL;
 827         struct lov_request_set *set;
 828         struct lov_obd *lov = &exp->exp_obd->u.lov;
 829         int rc = 0, i, shift;
 830         ENTRY;
 831
 832         OBD_ALLOC(set, sizeof(*set));
 833         if (set == NULL)
 834                 RETURN(-ENOMEM);
 835         lov_init_set(set);
 836
 837         set->set_exp = exp;
 838         set->set_oti = oti;
 839         set->set_oi = oinfo;
 840         set->set_oabufs = oa_bufs;
 841         OBD_ALLOC(set->set_pga, oa_bufs * sizeof(*set->set_pga));
 842         if (!set->set_pga)
 843                 GOTO(out, rc = -ENOMEM);
 844
 845         OBD_ALLOC(info, sizeof(*info) * oinfo->oi_md->lsm_stripe_count);
 846         if (!info)
 847                 GOTO(out, rc = -ENOMEM);
 848
 849         /* calculate the page count for each stripe */
 850         for (i = 0; i < oa_bufs; i++) {
 851                 int stripe = lov_stripe_number(oinfo->oi_md, pga[i].off);
 852                 info[stripe].count++;
 853         }
 854
 855         /* alloc and initialize lov request */
 856         shift = 0;
 857         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++){
 858                 struct lov_oinfo *loi = NULL;
 859                 struct lov_request *req;
 860
 861                 if (info[i].count == 0)
 862                         continue;
 863
 864                 loi = oinfo->oi_md->lsm_oinfo[i];
 865                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
 866                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
 867                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 868                         GOTO(out, rc = -EIO);
 869                 }
 870
 871                 OBD_ALLOC(req, sizeof(*req));
 872                 if (req == NULL)
 873                         GOTO(out, rc = -ENOMEM);
 874
 875                 OBDO_ALLOC(req->rq_oi.oi_oa);
 876                 if (req->rq_oi.oi_oa == NULL) {
 877                         OBD_FREE(req, sizeof(*req));
 878                         GOTO(out, rc = -ENOMEM);
 879                 }
 880
 881                 if (oinfo->oi_oa) {
 882                         memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
 883                                sizeof(*req->rq_oi.oi_oa));
 884                 }
 885                 req->rq_oi.oi_oa->o_id = loi->loi_id;
 886                 req->rq_oi.oi_oa->o_stripe_idx = i;
 887
 888                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 889                 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
 890                 if (req->rq_oi.oi_md == NULL) {
 891                         OBDO_FREE(req->rq_oi.oi_oa);
 892                         OBD_FREE(req, sizeof(*req));
 893                         GOTO(out, rc = -ENOMEM);
 894                 }
 895
 896                 req->rq_idx = loi->loi_ost_idx;
 897                 req->rq_stripe = i;
 898
 899                 /* XXX LOV STACKING */
 900                 req->rq_oi.oi_md->lsm_object_id = loi->loi_id;
 901                 req->rq_oi.oi_md->lsm_object_gr = oinfo->oi_md->lsm_object_gr;
 902                 req->rq_oabufs = info[i].count;
 903                 req->rq_pgaidx = shift;
 904                 shift += req->rq_oabufs;
 905
 906                 /* remember the index for sort brw_page array */
 907                 info[i].index = req->rq_pgaidx;
 908
 909                 req->rq_oi.oi_capa = oinfo->oi_capa;
 910
 911                 lov_set_add_req(req, set);
 912         }
 913         if (!set->set_count)
 914                 GOTO(out, rc = -EIO);
 915
 916         /* rotate & sort the brw_page array */
 917         for (i = 0; i < oa_bufs; i++) {
 918                 int stripe = lov_stripe_number(oinfo->oi_md, pga[i].off);
 919
 920                 shift = info[stripe].index + info[stripe].off;
 921                 LASSERT(shift < oa_bufs);
 922                 set->set_pga[shift] = pga[i];
 923                 lov_stripe_offset(oinfo->oi_md, pga[i].off, stripe,
 924                                   &set->set_pga[shift].off);
 925                 info[stripe].off++;
 926         }
 927 out:
 928         if (info)
 929                 OBD_FREE(info, sizeof(*info) * oinfo->oi_md->lsm_stripe_count);
 930
 931         if (rc == 0)
 932                 *reqset = set;
 933         else
 934                 lov_fini_brw_set(set);
 935
 936         RETURN(rc);
 937 }
 938
 939 int lov_fini_getattr_set(struct lov_request_set *set)
 940 {
 941         int rc = 0;
 942         ENTRY;
 943
 944         if (set == NULL)
 945                 RETURN(0);
 946         LASSERT(set->set_exp);
 947         if (set->set_completes)
 948                 rc = common_attr_done(set);
 949
 950         if (atomic_dec_and_test(&set->set_refcount))
 951                 lov_finish_set(set);
 952
 953         RETURN(rc);
 954 }
 955
 956 /* The callback for osc_getattr_async that finilizes a request info when a
 957  * response is recieved. */
 958 static int cb_getattr_update(void *cookie, int rc)
 959 {
 960         struct obd_info *oinfo = cookie;
 961         struct lov_request *lovreq;
 962         lovreq = container_of(oinfo, struct lov_request, rq_oi);
 963         return lov_update_common_set(lovreq->rq_rqset, lovreq, rc);
 964 }
 965
 966 int lov_prep_getattr_set(struct obd_export *exp, struct obd_info *oinfo,
 967                          struct lov_request_set **reqset)
 968 {
 969         struct lov_request_set *set;
 970         struct lov_obd *lov = &exp->exp_obd->u.lov;
 971         int rc = 0, i;
 972         ENTRY;
 973
 974         OBD_ALLOC(set, sizeof(*set));
 975         if (set == NULL)
 976                 RETURN(-ENOMEM);
 977         lov_init_set(set);
 978
 979         set->set_exp = exp;
 980         set->set_oi = oinfo;
 981
 982         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
 983                 struct lov_oinfo *loi;
 984                 struct lov_request *req;
 985
 986                 loi = oinfo->oi_md->lsm_oinfo[i];
 987                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
 988                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
 989                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 990                         continue;
 991                 }
 992
 993                 OBD_ALLOC(req, sizeof(*req));
 994                 if (req == NULL)
 995                         GOTO(out_set, rc = -ENOMEM);
 996
 997                 req->rq_stripe = i;
 998                 req->rq_idx = loi->loi_ost_idx;
 999
1000                 OBDO_ALLOC(req->rq_oi.oi_oa);
1001                 if (req->rq_oi.oi_oa == NULL) {
1002                         OBD_FREE(req, sizeof(*req));
1003                         GOTO(out_set, rc = -ENOMEM);
1004                 }
1005                 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
1006                        sizeof(*req->rq_oi.oi_oa));
1007                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1008                 req->rq_oi.oi_cb_up = cb_getattr_update;
1009                 req->rq_oi.oi_capa = oinfo->oi_capa;
1010                 req->rq_rqset = set;
1011
1012                 lov_set_add_req(req, set);
1013         }
1014         if (!set->set_count)
1015                 GOTO(out_set, rc = -EIO);
1016         *reqset = set;
1017         RETURN(rc);
1018 out_set:
1019         lov_fini_getattr_set(set);
1020         RETURN(rc);
1021 }
1022
1023 int lov_fini_destroy_set(struct lov_request_set *set)
1024 {
1025         ENTRY;
1026
1027         if (set == NULL)
1028                 RETURN(0);
1029         LASSERT(set->set_exp);
1030         if (set->set_completes) {
1031                 /* FIXME update qos data here */
1032         }
1033
1034         if (atomic_dec_and_test(&set->set_refcount))
1035                 lov_finish_set(set);
1036
1037         RETURN(0);
1038 }
1039
1040 int lov_prep_destroy_set(struct obd_export *exp, struct obd_info *oinfo,
1041                          struct obdo *src_oa, struct lov_stripe_md *lsm,
1042                          struct obd_trans_info *oti,
1043                          struct lov_request_set **reqset)
1044 {
1045         struct lov_request_set *set;
1046         struct lov_obd *lov = &exp->exp_obd->u.lov;
1047         int rc = 0, i;
1048         ENTRY;
1049
1050         OBD_ALLOC(set, sizeof(*set));
1051         if (set == NULL)
1052                 RETURN(-ENOMEM);
1053         lov_init_set(set);
1054
1055         set->set_exp = exp;
1056         set->set_oi = oinfo;
1057         set->set_oi->oi_md = lsm;
1058         set->set_oi->oi_oa = src_oa;
1059         set->set_oti = oti;
1060         if (oti != NULL && src_oa->o_valid & OBD_MD_FLCOOKIE)
1061                 set->set_cookies = oti->oti_logcookies;
1062
1063         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1064                 struct lov_oinfo *loi;
1065                 struct lov_request *req;
1066
1067                 loi = lsm->lsm_oinfo[i];
1068                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1069                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1070                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1071                         continue;
1072                 }
1073
1074                 OBD_ALLOC(req, sizeof(*req));
1075                 if (req == NULL)
1076                         GOTO(out_set, rc = -ENOMEM);
1077
1078                 req->rq_stripe = i;
1079                 req->rq_idx = loi->loi_ost_idx;
1080
1081                 OBDO_ALLOC(req->rq_oi.oi_oa);
1082                 if (req->rq_oi.oi_oa == NULL) {
1083                         OBD_FREE(req, sizeof(*req));
1084                         GOTO(out_set, rc = -ENOMEM);
1085                 }
1086                 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
1087                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1088                 lov_set_add_req(req, set);
1089         }
1090         if (!set->set_count)
1091                 GOTO(out_set, rc = -EIO);
1092         *reqset = set;
1093         RETURN(rc);
1094 out_set:
1095         lov_fini_destroy_set(set);
1096         RETURN(rc);
1097 }
1098
1099 int lov_fini_setattr_set(struct lov_request_set *set)
1100 {
1101         int rc = 0;
1102         ENTRY;
1103
1104         if (set == NULL)
1105                 RETURN(0);
1106         LASSERT(set->set_exp);
1107         if (set->set_completes) {
1108                 rc = common_attr_done(set);
1109                 /* FIXME update qos data here */
1110         }
1111
1112         if (atomic_dec_and_test(&set->set_refcount))
1113                 lov_finish_set(set);
1114         RETURN(rc);
1115 }
1116
1117 int lov_update_setattr_set(struct lov_request_set *set,
1118                            struct lov_request *req, int rc)
1119 {
1120         struct lov_obd *lov = &req->rq_rqset->set_exp->exp_obd->u.lov;
1121         struct lov_stripe_md *lsm = req->rq_rqset->set_oi->oi_md;
1122         ENTRY;
1123
1124         lov_update_set(set, req, rc);
1125
1126         /* grace error on inactive ost */
1127         if (rc && !(lov->lov_tgts[req->rq_idx] &&
1128                     lov->lov_tgts[req->rq_idx]->ltd_active))
1129                 rc = 0;
1130
1131         if (rc == 0) {
1132                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLCTIME)
1133                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_ctime =
1134                                 req->rq_oi.oi_oa->o_ctime;
1135                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLMTIME)
1136                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_mtime =
1137                                 req->rq_oi.oi_oa->o_mtime;
1138                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLATIME)
1139                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_atime =
1140                                 req->rq_oi.oi_oa->o_atime;
1141         }
1142
1143         RETURN(rc);
1144 }
1145
1146 /* The callback for osc_setattr_async that finilizes a request info when a
1147  * response is recieved. */
1148 static int cb_setattr_update(void *cookie, int rc)
1149 {
1150         struct obd_info *oinfo = cookie;
1151         struct lov_request *lovreq;
1152         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1153         return lov_update_setattr_set(lovreq->rq_rqset, lovreq, rc);
1154 }
1155
1156 int lov_prep_setattr_set(struct obd_export *exp, struct obd_info *oinfo,
1157                          struct obd_trans_info *oti,
1158                          struct lov_request_set **reqset)
1159 {
1160         struct lov_request_set *set;
1161         struct lov_obd *lov = &exp->exp_obd->u.lov;
1162         int rc = 0, i;
1163         ENTRY;
1164
1165         OBD_ALLOC(set, sizeof(*set));
1166         if (set == NULL)
1167                 RETURN(-ENOMEM);
1168         lov_init_set(set);
1169
1170         set->set_exp = exp;
1171         set->set_oti = oti;
1172         set->set_oi = oinfo;
1173         if (oti != NULL && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
1174                 set->set_cookies = oti->oti_logcookies;
1175
1176         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
1177                 struct lov_oinfo *loi = oinfo->oi_md->lsm_oinfo[i];
1178                 struct lov_request *req;
1179
1180                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1181                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1182                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1183                         continue;
1184                 }
1185
1186                 OBD_ALLOC(req, sizeof(*req));
1187                 if (req == NULL)
1188                         GOTO(out_set, rc = -ENOMEM);
1189                 req->rq_stripe = i;
1190                 req->rq_idx = loi->loi_ost_idx;
1191
1192                 OBDO_ALLOC(req->rq_oi.oi_oa);
1193                 if (req->rq_oi.oi_oa == NULL) {
1194                         OBD_FREE(req, sizeof(*req));
1195                         GOTO(out_set, rc = -ENOMEM);
1196                 }
1197                 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
1198                        sizeof(*req->rq_oi.oi_oa));
1199                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1200                 LASSERT(!(req->rq_oi.oi_oa->o_valid & OBD_MD_FLGROUP)
1201                                 || req->rq_oi.oi_oa->o_gr>0);
1202                 req->rq_oi.oi_oa->o_stripe_idx = i;
1203                 req->rq_oi.oi_cb_up = cb_setattr_update;
1204                 req->rq_oi.oi_capa = oinfo->oi_capa;
1205                 req->rq_rqset = set;
1206
1207                 if (oinfo->oi_oa->o_valid & OBD_MD_FLSIZE) {
1208                         int off = lov_stripe_offset(oinfo->oi_md,
1209                                                     oinfo->oi_oa->o_size, i,
1210                                                     &req->rq_oi.oi_oa->o_size);
1211
1212                         if (off < 0 && req->rq_oi.oi_oa->o_size)
1213                                 req->rq_oi.oi_oa->o_size--;
1214
1215                         CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1216                                i, req->rq_oi.oi_oa->o_size,
1217                                oinfo->oi_oa->o_size);
1218                 }
1219                 lov_set_add_req(req, set);
1220         }
1221         if (!set->set_count)
1222                 GOTO(out_set, rc = -EIO);
1223         *reqset = set;
1224         RETURN(rc);
1225 out_set:
1226         lov_fini_setattr_set(set);
1227         RETURN(rc);
1228 }
1229
1230 int lov_fini_punch_set(struct lov_request_set *set)
1231 {
1232         int rc = 0;
1233         ENTRY;
1234
1235         if (set == NULL)
1236                 RETURN(0);
1237         LASSERT(set->set_exp);
1238         if (set->set_completes) {
1239                 rc = -EIO;
1240                 /* FIXME update qos data here */
1241                 if (set->set_success)
1242                         rc = common_attr_done(set);
1243         }
1244
1245         if (atomic_dec_and_test(&set->set_refcount))
1246                 lov_finish_set(set);
1247
1248         RETURN(rc);
1249 }
1250
1251 int lov_update_punch_set(struct lov_request_set *set,
1252                          struct lov_request *req, int rc)
1253 {
1254         struct lov_obd *lov = &req->rq_rqset->set_exp->exp_obd->u.lov;
1255         struct lov_stripe_md *lsm = req->rq_rqset->set_oi->oi_md;
1256         ENTRY;
1257
1258         lov_update_set(set, req, rc);
1259
1260         /* grace error on inactive ost */
1261         if (rc && !lov->lov_tgts[req->rq_idx]->ltd_active)
1262                 rc = 0;
1263
1264         if (rc == 0) {
1265                 lov_stripe_lock(lsm);
1266                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLBLOCKS) {
1267                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_blocks =
1268                                 req->rq_oi.oi_oa->o_blocks;
1269                 }
1270
1271                 /* Do we need to update lvb_size here? It needn't because
1272                  * it have been done in ll_truncate(). -jay */
1273                 lov_stripe_unlock(lsm);
1274         }
1275
1276         RETURN(rc);
1277 }
1278
1279 /* The callback for osc_punch that finilizes a request info when a response
1280  * is recieved. */
1281 static int cb_update_punch(void *cookie, int rc)
1282 {
1283         struct obd_info *oinfo = cookie;
1284         struct lov_request *lovreq;
1285         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1286         return lov_update_punch_set(lovreq->rq_rqset, lovreq, rc);
1287 }
1288
1289 int lov_prep_punch_set(struct obd_export *exp, struct obd_info *oinfo,
1290                        struct obd_trans_info *oti,
1291                        struct lov_request_set **reqset)
1292 {
1293         struct lov_request_set *set;
1294         struct lov_obd *lov = &exp->exp_obd->u.lov;
1295         int rc = 0, i;
1296         ENTRY;
1297
1298         OBD_ALLOC(set, sizeof(*set));
1299         if (set == NULL)
1300                 RETURN(-ENOMEM);
1301         lov_init_set(set);
1302
1303         set->set_oi = oinfo;
1304         set->set_exp = exp;
1305
1306         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
1307                 struct lov_oinfo *loi = oinfo->oi_md->lsm_oinfo[i];
1308                 struct lov_request *req;
1309                 obd_off rs, re;
1310
1311                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1312                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1313                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1314                         continue;
1315                 }
1316
1317                 if (!lov_stripe_intersects(oinfo->oi_md, i,
1318                                            oinfo->oi_policy.l_extent.start,
1319                                            oinfo->oi_policy.l_extent.end,
1320                                            &rs, &re))
1321                         continue;
1322
1323                 OBD_ALLOC(req, sizeof(*req));
1324                 if (req == NULL)
1325                         GOTO(out_set, rc = -ENOMEM);
1326                 req->rq_stripe = i;
1327                 req->rq_idx = loi->loi_ost_idx;
1328
1329                 OBDO_ALLOC(req->rq_oi.oi_oa);
1330                 if (req->rq_oi.oi_oa == NULL) {
1331                         OBD_FREE(req, sizeof(*req));
1332                         GOTO(out_set, rc = -ENOMEM);
1333                 }
1334                 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
1335                        sizeof(*req->rq_oi.oi_oa));
1336                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1337                 req->rq_oi.oi_oa->o_gr = loi->loi_gr;
1338                 req->rq_oi.oi_oa->o_valid |= OBD_MD_FLGROUP;
1339
1340                 req->rq_oi.oi_oa->o_stripe_idx = i;
1341                 req->rq_oi.oi_cb_up = cb_update_punch;
1342                 req->rq_rqset = set;
1343
1344                 req->rq_oi.oi_policy.l_extent.start = rs;
1345                 req->rq_oi.oi_policy.l_extent.end = re;
1346                 req->rq_oi.oi_policy.l_extent.gid = -1;
1347
1348                 req->rq_oi.oi_capa = oinfo->oi_capa;
1349
1350                 lov_set_add_req(req, set);
1351         }
1352         if (!set->set_count)
1353                 GOTO(out_set, rc = -EIO);
1354         *reqset = set;
1355         RETURN(rc);
1356 out_set:
1357         lov_fini_punch_set(set);
1358         RETURN(rc);
1359 }
1360
1361 int lov_fini_sync_set(struct lov_request_set *set)
1362 {
1363         int rc = 0;
1364         ENTRY;
1365
1366         if (set == NULL)
1367                 RETURN(0);
1368         LASSERT(set->set_exp);
1369         if (set->set_completes) {
1370                 if (!set->set_success)
1371                         rc = -EIO;
1372                 /* FIXME update qos data here */
1373         }
1374
1375         if (atomic_dec_and_test(&set->set_refcount))
1376                 lov_finish_set(set);
1377
1378         RETURN(rc);
1379 }
1380
1381 int lov_prep_sync_set(struct obd_export *exp, struct obd_info *oinfo,
1382                       struct obdo *src_oa, struct lov_stripe_md *lsm,
1383                       obd_off start, obd_off end,
1384                       struct lov_request_set **reqset)
1385 {
1386         struct lov_request_set *set;
1387         struct lov_obd *lov = &exp->exp_obd->u.lov;
1388         int rc = 0, i;
1389         ENTRY;
1390
1391         OBD_ALLOC(set, sizeof(*set));
1392         if (set == NULL)
1393                 RETURN(-ENOMEM);
1394         lov_init_set(set);
1395
1396         set->set_exp = exp;
1397         set->set_oi = oinfo;
1398         set->set_oi->oi_md = lsm;
1399         set->set_oi->oi_oa = src_oa;
1400
1401         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1402                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
1403                 struct lov_request *req;
1404                 obd_off rs, re;
1405
1406                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1407                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1408                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1409                         continue;
1410                 }
1411
1412                 if (!lov_stripe_intersects(lsm, i, start, end, &rs, &re))
1413                         continue;
1414
1415                 OBD_ALLOC(req, sizeof(*req));
1416                 if (req == NULL)
1417                         GOTO(out_set, rc = -ENOMEM);
1418                 req->rq_stripe = i;
1419                 req->rq_idx = loi->loi_ost_idx;
1420
1421                 OBDO_ALLOC(req->rq_oi.oi_oa);
1422                 if (req->rq_oi.oi_oa == NULL) {
1423                         OBD_FREE(req, sizeof(*req));
1424                         GOTO(out_set, rc = -ENOMEM);
1425                 }
1426                 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
1427                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1428                 req->rq_oi.oi_oa->o_stripe_idx = i;
1429
1430                 req->rq_oi.oi_policy.l_extent.start = rs;
1431                 req->rq_oi.oi_policy.l_extent.end = re;
1432                 req->rq_oi.oi_policy.l_extent.gid = -1;
1433
1434                 lov_set_add_req(req, set);
1435         }
1436         if (!set->set_count)
1437                 GOTO(out_set, rc = -EIO);
1438         *reqset = set;
1439         RETURN(rc);
1440 out_set:
1441         lov_fini_sync_set(set);
1442         RETURN(rc);
1443 }
1444
1445 #define LOV_U64_MAX ((__u64)~0ULL)
1446 #define LOV_SUM_MAX(tot, add)                                           \
1447         do {                                                            \
1448                 if ((tot) + (add) < (tot))                              \
1449                         (tot) = LOV_U64_MAX;                            \
1450                 else                                                    \
1451                         (tot) += (add);                                 \
1452         } while(0)
1453
1454 int lov_fini_statfs(struct obd_device *obd, struct obd_statfs *osfs,int success)
1455 {
1456         ENTRY;
1457
1458         if (success) {
1459                 __u32 expected_stripes = lov_get_stripecnt(&obd->u.lov, 0);
1460
1461                 if (osfs->os_files != LOV_U64_MAX)
1462                         do_div(osfs->os_files, expected_stripes);
1463                 if (osfs->os_ffree != LOV_U64_MAX)
1464                         do_div(osfs->os_ffree, expected_stripes);
1465
1466                 spin_lock(&obd->obd_osfs_lock);
1467                 memcpy(&obd->obd_osfs, osfs, sizeof(*osfs));
1468                 obd->obd_osfs_age = cfs_time_current_64();
1469                 spin_unlock(&obd->obd_osfs_lock);
1470                 RETURN(0);
1471         }
1472
1473         RETURN(-EIO);
1474 }
1475
1476 int lov_fini_statfs_set(struct lov_request_set *set)
1477 {
1478         int rc = 0;
1479         ENTRY;
1480
1481         if (set == NULL)
1482                 RETURN(0);
1483
1484         if (set->set_completes) {
1485                 rc = lov_fini_statfs(set->set_obd, set->set_oi->oi_osfs,
1486                                      set->set_success);
1487         }
1488
1489         if (atomic_dec_and_test(&set->set_refcount))
1490                 lov_finish_set(set);
1491
1492         RETURN(rc);
1493 }
1494
1495 void lov_update_statfs(struct obd_statfs *osfs, struct obd_statfs *lov_sfs,
1496                        int success)
1497 {
1498         int shift = 0, quit = 0;
1499         __u64 tmp;
1500
1501         if (success == 0) {
1502                 memcpy(osfs, lov_sfs, sizeof(*lov_sfs));
1503         } else {
1504                 if (osfs->os_bsize != lov_sfs->os_bsize) {
1505                         /* assume all block sizes are always powers of 2 */
1506                         /* get the bits difference */
1507                         tmp = osfs->os_bsize | lov_sfs->os_bsize;
1508                         for (shift = 0; shift <= 64; ++shift) {
1509                                 if (tmp & 1) {
1510                                         if (quit)
1511                                                 break;
1512                                         else
1513                                                 quit = 1;
1514                                         shift = 0;
1515                                 }
1516                                 tmp >>= 1;
1517                         }
1518                 }
1519
1520                 if (osfs->os_bsize < lov_sfs->os_bsize) {
1521                         osfs->os_bsize = lov_sfs->os_bsize;
1522
1523                         osfs->os_bfree  >>= shift;
1524                         osfs->os_bavail >>= shift;
1525                         osfs->os_blocks >>= shift;
1526                 } else if (shift != 0) {
1527                         lov_sfs->os_bfree  >>= shift;
1528                         lov_sfs->os_bavail >>= shift;
1529                         lov_sfs->os_blocks >>= shift;
1530                 }
1531 #ifdef MIN_DF
1532                 /* Sandia requested that df (and so, statfs) only
1533                    returned minimal available space on
1534                    a single OST, so people would be able to
1535                    write this much data guaranteed. */
1536                 if (osfs->os_bavail > lov_sfs->os_bavail) {
1537                         /* Presumably if new bavail is smaller,
1538                            new bfree is bigger as well */
1539                         osfs->os_bfree = lov_sfs->os_bfree;
1540                         osfs->os_bavail = lov_sfs->os_bavail;
1541                 }
1542 #else
1543                 osfs->os_bfree += lov_sfs->os_bfree;
1544                 osfs->os_bavail += lov_sfs->os_bavail;
1545 #endif
1546                 osfs->os_blocks += lov_sfs->os_blocks;
1547                 /* XXX not sure about this one - depends on policy.
1548                  *   - could be minimum if we always stripe on all OBDs
1549                  *     (but that would be wrong for any other policy,
1550                  *     if one of the OBDs has no more objects left)
1551                  *   - could be sum if we stripe whole objects
1552                  *   - could be average, just to give a nice number
1553                  *
1554                  * To give a "reasonable" (if not wholly accurate)
1555                  * number, we divide the total number of free objects
1556                  * by expected stripe count (watch out for overflow).
1557                  */
1558                 LOV_SUM_MAX(osfs->os_files, lov_sfs->os_files);
1559                 LOV_SUM_MAX(osfs->os_ffree, lov_sfs->os_ffree);
1560         }
1561 }
1562
1563 /* The callback for osc_statfs_async that finilizes a request info when a
1564  * response is recieved. */
1565 static int cb_statfs_update(void *cookie, int rc)
1566 {
1567         struct obd_info *oinfo = cookie;
1568         struct lov_request *lovreq;
1569         struct obd_statfs *osfs, *lov_sfs;
1570         struct obd_device *obd;
1571         struct lov_obd *lov;
1572         int success;
1573         ENTRY;
1574
1575         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1576         lov = &lovreq->rq_rqset->set_obd->u.lov;
1577         obd = class_exp2obd(lov->lov_tgts[lovreq->rq_idx]->ltd_exp);
1578
1579         osfs = lovreq->rq_rqset->set_oi->oi_osfs;
1580         lov_sfs = oinfo->oi_osfs;
1581
1582         success = lovreq->rq_rqset->set_success;
1583
1584         /* XXX: the same is done in lov_update_common_set, however
1585            lovset->set_exp is not initialized. */
1586         lov_update_set(lovreq->rq_rqset, lovreq, rc);
1587         if (rc) {
1588                 if (rc && !(lov->lov_tgts[lovreq->rq_idx] &&
1589                             lov->lov_tgts[lovreq->rq_idx]->ltd_active))
1590                         rc = 0;
1591                 RETURN(rc);
1592         }
1593
1594         spin_lock(&obd->obd_osfs_lock);
1595         memcpy(&obd->obd_osfs, lov_sfs, sizeof(*lov_sfs));
1596         if ((oinfo->oi_flags & OBD_STATFS_FROM_CACHE) == 0)
1597                 obd->obd_osfs_age = cfs_time_current_64();
1598         spin_unlock(&obd->obd_osfs_lock);
1599
1600         lov_update_statfs(osfs, lov_sfs, success);
1601         qos_update(lov);
1602
1603         RETURN(0);
1604 }
1605
1606 int lov_prep_statfs_set(struct obd_device *obd, struct obd_info *oinfo,
1607                         struct lov_request_set **reqset)
1608 {
1609         struct lov_request_set *set;
1610         struct lov_obd *lov = &obd->u.lov;
1611         int rc = 0, i;
1612         ENTRY;
1613
1614         OBD_ALLOC(set, sizeof(*set));
1615         if (set == NULL)
1616                 RETURN(-ENOMEM);
1617         lov_init_set(set);
1618
1619         set->set_obd = obd;
1620         set->set_oi = oinfo;
1621
1622         /* We only get block data from the OBD */
1623         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1624                 struct lov_request *req;
1625
1626                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active) {
1627                         CDEBUG(D_HA, "lov idx %d inactive\n", i);
1628                         continue;
1629                 }
1630
1631                 OBD_ALLOC(req, sizeof(*req));
1632                 if (req == NULL)
1633                         GOTO(out_set, rc = -ENOMEM);
1634
1635                 OBD_ALLOC(req->rq_oi.oi_osfs, sizeof(*req->rq_oi.oi_osfs));
1636                 if (req->rq_oi.oi_osfs == NULL) {
1637                         OBD_FREE(req, sizeof(*req));
1638                         GOTO(out_set, rc = -ENOMEM);
1639                 }
1640
1641                 req->rq_idx = i;
1642                 req->rq_oi.oi_cb_up = cb_statfs_update;
1643                 req->rq_oi.oi_flags = oinfo->oi_flags;
1644                 req->rq_rqset = set;
1645
1646                 lov_set_add_req(req, set);
1647         }
1648         if (!set->set_count)
1649                 GOTO(out_set, rc = -EIO);
1650         *reqset = set;
1651         RETURN(rc);
1652 out_set:
1653         lov_fini_statfs_set(set);
1654         RETURN(rc);
1655 }